<p><strong>TLDR:</strong> QCon London 2026 surfaced some genuinely interesting infrastructure work, with Tansu — a stateless, Rust-based Kafka-compatible broker — being the standout. OpenAI quietly shipped a container pool to the Responses API that cuts agent spin-up time by roughly 10x. The microservices-vs-monolith debate got another honest treatment, and Morgan Stanley is rethinking its entire API program around MCP.</p>
<hr>
<h2>Tansu: a Rust-based Kafka broker that fits in 20MB</h2>
<p>Peter Morgan presented Tansu at QCon London — a solo two-year project that replaces Kafka’s internal replication and state with pluggable external storage backends like S3, SQLite, and Postgres. Brokers are stateless and leaderless, use 20MB of RAM, and restart in about 10ms. The Postgres integration is particularly interesting: it uses the <code>COPY FROM</code> protocol for high-throughput ingestion and eliminates the transactional outbox pattern entirely. Direct writes to Iceberg, Delta Lake, and Parquet via “sink topics” are also supported. SSL, ACLs, and S3 compaction are still missing, so it’s not production-ready for most teams yet — but the architecture is worth watching.</p>
<h2>OpenAI adds container pooling to the Responses API</h2>
<p>OpenAI shipped a container pool for the Responses API, letting agent workflow components like the code interpreter and shell reuse warm infrastructure instead of spinning up fresh containers each session. The claimed improvement is around 10x faster startup. For anyone building agentic pipelines where cold-start latency is a real problem, this is a meaningful change — not a marketing number.</p>
<h2>Microservices vs. modular monolith, argued honestly</h2>
<p>A post making the rounds today makes the case that most teams reach for microservices to solve problems that are actually caused by poor module boundaries, missing tests, or database bottlenecks. The real costs — distributed transaction complexity, observability overhead, broken local dev experience — rarely get accounted for upfront. The modular monolith gets a fair hearing as the overlooked middle ground. Four questions are offered as a decision filter: independent deployment needs, divergent scaling profiles, observability already in place, and team bandwidth to own the boundary end-to-end.</p>
<h2>Morgan Stanley rethinks its API program for MCP</h2>
<p>At QCon London, Morgan Stanley presented how it’s adapting its API strategy around the Model Context Protocol. A major financial institution publicly rethinking API design for AI agent integration patterns is a signal worth noting — enterprise API programs that were built around human-facing clients are going to need structural changes as agents become primary consumers.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Netflix ontology-driven observability:</strong> A QCon London talk covered how Netflix uses a knowledge graph to model relationships between services, infrastructure, and telemetry — giving semantic context to observability signals for root cause analysis at scale.</li>
<li><strong>OpenTelemetry for queue bottlenecks:</strong> Mark Silvester’s QCon session focused on using OpenTelemetry specifically to diagnose performance issues in message queues and async pipelines — a narrower, more practical angle than the usual OTel overview.</li>
<li><strong>Kafka + TypeScript + Protobuf walkthrough:</strong> A step-by-step guide covers running Kafka locally with Docker, using the <code>@platformatic/kafka</code> client, encoding with Protobuf for a shared schema contract, and integrating event publishing into a gRPC-based product service.</li>
<li><strong>RunPod Flash for serverless GPU functions:</strong> RunPod’s new SDK lets you deploy GPU functions directly from Python using decorators — no Docker image, no registry. Each function gets its own serverless endpoint and scales independently, which makes sense for on-demand AI workloads with uneven traffic.</li>
<li><strong>tRPC vs. Elysia:</strong> A developer evaluated migrating from tRPC to Elysia to make their API more “agent friendly” and found no reason to switch — tRPC with plugins covers the same ground.</li>
<li><strong>Agentic platform engineering pattern:</strong> A proposed architecture treats AI agent configuration as versioned infrastructure across three repos: agent-library, agent-setup, and resource-catalog. Layered Markdown files scoped to filesystem directories keep token usage down; full environment rebuild targets under 5 minutes.</li>
<li><strong>GraphQL introspection still left on in production:</strong> A practical scraping guide notes that introspection is frequently enabled in production GraphQL APIs, often exposing more fields than the UI surfaces — a reminder to audit this if you’re running GraphQL publicly.</li>
</ul>


Server-side development, system design, API patterns, and distributed architecture.

Backend Digest

A roundup of infrastructure highlights from QCon London 2026 and recent engineering news. Tansu is a solo-built, stateless Rust-based Kafka-compatible broker using pluggable backends like S3, SQLite, and Postgres — fitting in 20MB RAM with 10ms restarts, though not yet production-ready. OpenAI added container pooling to its Responses API, cutting agent cold-start times by ~10x. The microservices-vs-modular-monolith debate gets an honest treatment, arguing most teams adopt microservices to mask poor module boundaries rather than genuine scaling needs. Morgan Stanley is restructuring its API program around the Model Context Protocol for AI agent integration. Additional items cover Netflix's ontology-driven observability, OpenTelemetry for queue diagnostics, Kafka with TypeScript and Protobuf, RunPod's serverless GPU SDK, tRPC vs. Elysia, agentic platform engineering patterns, and GraphQL introspection risks in production.

Tansu rewrites Kafka in Rust, OpenAI warms containers 10x faster